Small area estimation using incomplete auxiliary information
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Small area estimation using incomplete auxiliary information".
Jane: Here's the gist of the paper. Small area estimation using incomplete auxiliary information proposes a twostep approach that uses design-based model calibration with non-probability data to obtain approximately design-unbiased domain totals, which are then used in a Fay–Herriot area-level model to produce empirical best linear unbiased predictors The Proposed Methodology The paper develops a methodology that combines probability sample data with incomplete auxiliary information, following the model-calibration framework of Wu and Sitter (2001) and adapting ideas from Kim and Tam (2021) on data integration This approach is motivated by three applications from official statistics on enterprises in Lithuania, where closely related value-added tax (VAT) data are available only for part of the population, annual investment in tangible assets with a cut-off sample providing partial information, and quarterly job vacancies obtained from non-probability web scraping The framework is best understood in terms of Type IV classification by Tam and Holmberg (2020), where proxy measurements are available only for a subset of the population through a non-probability source and can be incorporated as auxiliary information Model Calibration Framework The model-calibration step involves determining calibration weights wi for i in Am so that they remain close to the baseline weights di under a chi-square distance while satisfying post-stratification constraints that align the probability sample Am with the nonprobability sample B The calibration is performed by imposing domain-specific constraints, including matching population sizes of missing-data and big-data strata within each domain and aligning the domain total of model-based predictions The resulting model-calibration estimator of the domain total is tˆMCm defined by X i∈Am wiyi, m = 1,..., M Variance Estimation and Area-Level Modeling The variance of the model-calibration estimator tˆMCm can be robustly estimated by resampling using the design-based bootstrap of Rao et al.
Tom: (1992) rather than relying on linearization In area-level modeling, log-transformation of direct domain estimators helps stabilize variances and satisfies model assumptions for the Fay–Herriot (FH) model The empirical best linear unbiased predictor (EBLUP) is then produced by treating the sampling variances as known when estimating the parameters β and σ2u, borrowing strength across domains through covariates based on completely known auxiliary data Empirical Results and Gains Empirical comparisons show consistent improvements in domain-level precision over direct estimation and over a Fay–Herriot benchmark that directly incorporates the proxy information as an error-prone covariate In the monthly turnover application, incorporating the VAT-based proxy through the unit-level model-calibration step leads to a clear and systematic precision gain, with the average number of domains classified as good increasing from 50.8 to 56.2 The largest relative reductions occur in the least precise domains under direct estimation, resulting in fewer domains being classified as unreliable Comparative Summary The main precision gains are achieved already at the model-calibration stage, with the FH step typically providing a secondary tightening of the distribution The magnitude of the gains varies across applications and tends to increase with both the relative size of the non-probability source and the strength of the unit-level relationship between y⋆ and y The Ybarra–Lohr benchmark exhibits more heterogeneous performance across applications, remaining relatively close to the direct estimator for turnover and vacancies whereas it is more beneficial for investment by reducing the remaining unreliable domains Overall, the unit-level prediction followed by model calibration appears to be the most consistently effective mechanism for improving domain precision across the considered settings The proposed two-step framework exploits incomplete auxiliary information without modeling the selection mechanism of the non-probability sample, yielding approximately design-unbiased estimators of domain totals with design-based uncertainty assessment The approach is flexible in the choice of working model linking y and y⋆ and can accommodate linear, nonlinear, or nonparametric relationships Empirical results indicate that the model calibration (MC) step accounts for most of the improvement in domain-level precision, while modeling the MC totals with the FH model typically provides an additional, smaller gain once the MC totals and their variance estimates are available The achievable gains depend on the quality of the proxy and the adequacy of the measurement-error working model References Battese, G.
Jane: E., Harter, R.
Tom: M., and Fuller, W.
Title and authors: Tom: We're starting with "Small area estimation using incomplete auxiliary information." The title itself tells you immediately that this is about estimating totals in small geographic areas when you don't have all the perfect data.
Jane: It sounds technical, but I think it boils down to this: we have some good survey numbers, but we also have these other big datasets that are usually partial or come from non-probability sources.
Lu: That’s the challenge they’re addressing. They want to show how you can use that incomplete auxiliary information—things like administrative registers or online job ads—to get better totals than you could with just the primary survey sample alone.
Meng: So, what are these authors trying to prove? Are they just showing a clever way to combine data sources, or is there something more fundamental about how we should handle selection bias?
Tom: They’re proposing a twostep approach that uses design-based model calibration first to get approximately design-unbiased domain totals, and then they use those totals in an area-level Fay Herriot model with exactly known covariates.
Jane: That sounds like a very specific workflow. So, what’s the big implication for someone listening who isn't a statistician? It means we can make better estimates when our data is messy.
Lu: The implication is that you don't have to assume every piece of information is perfect or comes from the right sampling method; you can use it as an input, provided you handle its uncertainty correctly.
Meng: But I’m still stuck on that selection mechanism thing. If we can’t model how the non-probability data was picked, isn't that a massive blind spot for any robust system?
Tom: They explicitly state they don't try to model the selection mechanism of the non-probability sample in this first step, which is a key difference from other methods.
Jane: So it’s about accepting that uncertainty there and just focusing on how to use that noisy proxy data effectively within a calibration framework.
Lu: Right. They develop a design-based method that uses incomplete auxiliary information to improve domain-level estimation, and then the resulting estimates can be modeled within small area estimation frameworks.
Meng: So they’re saying this framework is flexible enough to handle different types of data structure, like Type IV data which involves both probability and non-probability samples simultaneously.
Tom: They even classify the data sources into four types: Type I, Type II, Type III, and Type IV based on what kind of auxiliary variables you have alongside your study variable y.
Jane: That classification helps us understand when this method is most applicable. It’s not a one-size-fits-all fix for every messy data problem out there.
The paper's summary: Tom: So, let’s go over the core summary of "Small area estimation using incomplete auxiliary information" again. The paper sets up this two-step process where you use design-based model calibration with non-probability data to get domain totals, which are then used in a Fay Herriot area-level model to produce empirical best linear unbiased predictors.
Jane: It sounds like the main takeaway is that combining probability sample data with incomplete auxiliary information allows us to obtain roughly design-unbiased domain totals that can then feed into a more powerful area-level model.
Lu: Exactly. The first step is using a unit-level measurement error working model on the link between the survey and external source, and then applying domain constraints to enforce alignment with the probability sample.
Meng: So, in simple terms, they are essentially creating high-quality totals by correcting for measurement errors before moving on to the final area modeling.
Tom: Right. They’re using that calibrated estimator of the domain total, t hat MC m, and then they use its variance estimates to feed into a Fay Herriot model with exactly known covariates for an empirical best linear unbiased predictor.
Jane: So what does that mean for someone listening? It means we move from just having potentially biased totals to having totals that have a quantified idea of how much uncertainty they carry, which is better than just trusting the raw numbers blindly.
Lu: They are using design-based bootstrap methods to estimate the variance of that calibration estimator, t hat MC m, instead of relying on linearization for variance estimation.
Meng: That sounds like a rigorous way to handle the uncertainty. It’s not just a quick fix; they are quantifying it systematically across different domains.
Tom: They also emphasize that log-transformation of direct domain estimators helps stabilize variances and satisfies the assumptions needed for the Fay Herriot model, which is important for area modeling.
Jane: So, if I’m using this instead of just a direct estimator, am I getting a benefit from stabilizing the variance and making sure my final model assumptions are met?
The paper's improvements: Tom: Now let’s talk about what they actually claim as improvements in this methodology. They suggest several ways to make this approach better than simpler methods.
Lu: One big improvement is that the model calibration for domain totals can be used to obtain approximately design-unbiased domain totals by fitting a unit-level measurement error working model on the linked overlap between the probability survey and the external source.
Jane: That sounds like they’re using that link between the two data sources as a way to correct for noise in one of them.
Tom: Exactly. And this allows AI systems to produce more reliable estimates of small area totals when auxiliary information is incomplete or noisy, as demonstrated in applications like monthly turnover.
Meng: That’s useful, but what about the other improvements? Are they just making the initial total estimate better, or are they changing how we do the final prediction too?
Lu: The framework also enables valid inference without ignoring assumptions on the selection mechanism of non-probability samples, contrasting with traditional methods that rely on an unverifiable missing at random assumption.
Jane: So this is a big deal for researchers because it means you can use those large, non-probability data sources without having to model the complex selection process of those sources.
Tom: And they also say the approach allows for substantially more general relationships between the observed proxies and your study variables than just simple linear models.
Meng: That means we can handle nonlinear stuff, right? Not just straight lines. I need that flexibility when things aren't perfectly linear in reality.
Lu: Yes, it accommodates linear, nonlinear, or nonparametric relationships in the measurement error model. It lets you integrate diverse auxiliary information types into estimation tasks where simple linear assumptions are insufficient.
Jane: That’s really helpful for real-world data because real relationships rarely fit a simple straight line perfectly.
Conclusion: Tom: Alright, let's wrap up with the conclusion of "Small area estimation using incomplete auxiliary information." They summarize the main points they found about their framework.
Jane: So, they are concluding that this two-step framework exploits incomplete auxiliary information without modeling the selection mechanism of the non-probability sample to yield approximately design-unbiased estimators of domain totals with design-based uncertainty assessment.
Lu: The paper shows that the model calibration step accounts for most of the improvement in domain-level precision, while modeling those calibrated totals with the Fay Herriot model typically provides an additional, smaller gain once those MC totals and their variance estimates are available.
Meng: So the main takeaway is that we should focus on getting that initial calibration right first.
Tom: That’s right. The achievable gains depend on the quality of the proxy and how adequate your measurement error working model is.
Jane: It seems like this method provides a very practical and powerful mechanism for improving domain precision across various settings by integrating different data types in a thoughtful way.
Lu: The paper demonstrates that this two-step framework can be flexible enough to accommodate diverse relationships between variables in the measurement error model.
Meng: It’s about using incomplete auxiliary information smartly, which is what matters for getting reliable results when your data isn't perfect.
Donatas šlevska, řeva BurakauskaitŨe Andrius Чiginas
Vilnius University, Faculty of Mathematics and Informatics, Institute of Data Science and Digital Technologies
stat.ME
Submitted: 2026-02-13
Updated: 2026-10-30
Comments: 37 pages, 3 figures, including supplementary material. Revised methodology and expanded empirical analysis; added a design-based simulation study
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 72/100
The gist: Small area estimation using incomplete auxiliary information proposes a twostep approach that uses design-based model calibration with non-probability data to obtain approximately design-unbiased…
Key concepts
- Model Calibration Framework
- This step adjusts weights for the probability sample so they align with constraints imposed by non-probability data. It ensures that the estimated domain totals are as unbiased as possible while respecting known population sizes and matching domain totals from model predictions.
- Fay–Herriot (FH) Model
- This area-level model uses log-transformed domain totals to produce empirical best linear unbiased predictors (EBLUPs). It allows predictions for unknown areas by borrowing statistical strength across domains using covariates derived from completely known auxiliary data, stabilizing the estimates.
- Design-Based Bootstrap
- This is a resampling technique used to robustly estimate the variance of the model calibration estimator. Instead of relying on simple linear approximations, it uses resampling methods to account for complex sampling designs and provide reliable uncertainty estimates.
- Incomplete Auxiliary Information
- This refers to using data that is available only for a subset of the population or from non-probability sources, like web scraping. The method integrates this noisy or partial information into the estimation process without needing to know the exact selection mechanism of that source.
Terminology
Summary
Small area estimation using incomplete auxiliary information proposes a twostep approach that uses design-based model calibration with non-probability data to obtain approximately design-unbiased domain totals, which are then used in a Fay–Herriot area-level model to produce empirical best linear unbiased predictors
The Proposed Methodology
The paper develops a methodology that combines probability sample data with incomplete auxiliary information, following the model-calibration framework of Wu and Sitter (2001) and adapting ideas from Kim and Tam (2021) on data integration This approach is motivated by three applications from official statistics on enterprises in Lithuania, where closely related value-added tax (VAT) data are available only for part of the population, annual investment in tangible assets with a cut-off sample providing partial information, and quarterly job vacancies obtained from non-probability web scraping The framework is best understood in terms of Type IV classification by Tam and Holmberg (2020), where proxy measurements are available only for a subset of the population through a non-probability source and can be incorporated as auxiliary information
Model Calibration Framework
The model-calibration step involves determining calibration weights wi for i in Am so that they remain close to the baseline weights di under a chi-square distance while satisfying post-stratification constraints that align the probability sample Am with the nonprobability sample B The calibration is performed by imposing domain-specific constraints, including matching population sizes of missing-data and big-data strata within each domain and aligning the domain total of model-based predictions The resulting model-calibration estimator of the domain total is tˆMCm defined by X i∈Am wiyi, m = 1,..., M
Variance Estimation and Area-Level Modeling
The variance of the model-calibration estimator tˆMCm can be robustly estimated by resampling using the design-based bootstrap of Rao et al. (1992) rather than relying on linearization In area-level modeling, log-transformation of direct domain estimators helps stabilize variances and satisfies model assumptions for the Fay–Herriot (FH) model The empirical best linear unbiased predictor (EBLUP) is then produced by treating the sampling variances as known when estimating the parameters β and σ2u, borrowing strength across domains through covariates based on completely known auxiliary data
Empirical Results and Gains
Empirical comparisons show consistent improvements in domain-level precision over direct estimation and over a Fay–Herriot benchmark that directly incorporates the proxy information as an error-prone covariate In the monthly turnover application, incorporating the VAT-based proxy through the unit-level model-calibration step leads to a clear and systematic precision gain, with the average number of domains classified as good increasing from 50.8 to 56.2 The largest relative reductions occur in the least precise domains under direct estimation, resulting in fewer domains being classified as unreliable
Comparative Summary
The main precision gains are achieved already at the model-calibration stage, with the FH step typically providing a secondary tightening of the distribution The magnitude of the gains varies across applications and tends to increase with both the relative size of the non-probability source and the strength of the unit-level relationship between y⋆ and y The Ybarra–Lohr benchmark exhibits more heterogeneous performance across applications, remaining relatively close to the direct estimator for turnover and vacancies whereas it is more beneficial for investment by reducing the remaining unreliable domains Overall, the unit-level prediction followed by model calibration appears to be the most consistently effective mechanism for improving domain precision across the considered settings
Conclusions
The proposed two-step framework exploits incomplete auxiliary information without modeling the selection mechanism of the non-probability sample, yielding approximately design-unbiased estimators of domain totals with design-based uncertainty assessment The approach is flexible in the choice of working model linking y and y⋆ and can accommodate linear, nonlinear, or nonparametric relationships Empirical results indicate that the model calibration (MC) step accounts for most of the improvement in domain-level precision, while modeling the MC totals with the FH model typically provides an additional, smaller gain once the MC totals and their variance estimates are available The achievable gains depend on the quality of the proxy and the adequacy of the measurement-error working model
References
Battese, G. E., Harter, R. M., and Fuller, W. A. (1988). An error-components model for prediction of county crop areas using survey and satellite data
Booth, J. G., Butler, R. W., and Hall, P. (1994). Bootstrap methods for finite populations
Chen, Y., Li, P., and Wu, C. (2020). Doubly robust inference with nonprobability survey samples
Deville, J.-C. and Särndal, C. p. 3 , xip)
⊤, p ≥ 1,
are available for nearly all population units from administrative registers, censuses, or other
external sources. Such information can be used to improve the direct estimators, such as
(1), through the standard calibration approach of Deville and Särndal (1992) or generalized
regression (GREG) estimation, leading to approximately design-unbiased estimators with
smaller variances at the domain level. , xip)
⊤, p ≥ 1,
are available for nearly all population units from administrative registers, censuses, or other
external sources. Such information can be used to improve the direct estimators, such as
(1), through the standard calibration approach of Deville and Särndal (1992) or generalized
regression (GREG) estimation, leading to approximately design-unbiased estimators with
smaller variances at the domain level.Where the paper says it
Where the paper says it
Elliott, M. and Valliant, R. p. 1 The setup is also closely related to the Type III data integration of Kim
and Tam (2021), which applies design-based estimation methods; however, our formulation
extends beyond their linear measurement-error model. For comparison, Type II approaches –
mass imputation, propensity-based inverse probability weighting (IPW), and doubly-robust
adjustments using a supplementary probability sample (Elliott and Valliant, 2017; Chen
et al. The setup is also closely related to the Type III data integration of Kim
and Tam (2021), which applies design-based estimation methods; however, our formulation
extends beyond their linear measurement-error model. For comparison, Type II approaches –
mass imputation, propensity-based inverse probability weighting (IPW), and doubly-robust
adjustments using a supplementary probability sample (Elliott and Valliant, 2017; Chen
et al.Where the paper says it
Where the paper says it
Fay, R. E. and Herriot, R. p. 0 A unit-level measurement-error working
model is fitted on the linked overlap between the probability survey and the external
source, and its predictions are incorporated through domain-specific model-calibration
constraints to obtain approximately design-unbiased domain totals. These totals and
their variance estimates are then used in a Fay–Herriot area-level model with exactly
known covariates to produce empirical best linear unbiased predictors. A unit-level measurement-error working
model is fitted on the linked overlap between the probability survey and the external
source, and its predictions are incorporated through domain-specific model-calibration
constraints to obtain approximately design-unbiased domain totals. These totals and
their variance estimates are then used in a Fay–Herriot area-level model with exactly
known covariates to produce empirical best linear unbiased predictors.Where the paper says it
Where the paper says it
Harmening, S., Kreutzmann, A.-K., Schmidt, S., Salvati, N., and Schmid, T. (2023). A framework for producing small area estimates based on area-level models in R
Jiang, J., Lahiri, P., and Wan, S.-M. (2002). A unified jackknife theory for empirical best prediction with M-estimation
Kim, J., Park, S., Chen, Y., and Wu, C. (2021). Combining non-probability and probability survey samples through mass imputation
Kim, J.-K. and Tam, S. p. 1 This setting makes classical calibration or model-assisted estimation methods inapplicable
in their standard form. From the perspective of non-probability sampling, our setting is best understood in terms
of the four-type classification summarized by Tam and Holmberg (2020). This setting makes classical calibration or model-assisted estimation methods inapplicable
in their standard form. From the perspective of non-probability sampling, our setting is best understood in terms
of the four-type classification summarized by Tam and Holmberg (2020).Where the paper says it
Where the paper says it
Liu, A.-C., Scholtus, S., and de Waal, T. (2023). Correcting selection bias in big data by pseudo-weighting
Medous, E., Goga, C., Ruiz-Gazen, A., Beaumont, J.-F., Dessertaine, A., and Puech, P. (2024). QR prediction for statistical data integration
Rao, J. N. K. and Molina, I. p. 2 (1988)) and area-level models (e. , Fay and Herriot (1979)); a comprehensive
account is given by Rao and Molina (2015). (1988)) and area-level models (e. , Fay and Herriot (1979)); a comprehensive
account is given by Rao and Molina (2015).Where the paper says it
Where the paper says it
Rao, J. N. K., Wu, C., and Yue, K. p. 5 In our applications considered in Section 3, the survey data are obtained from stratified
simple random samples without replacement, with strata defined by enterprise characteristics
such as economic activity and size class. The Rao–Wu–Yue bootstrap was developed for
general stratified multistage designs and can be directly applied to stratified simple random
sampling. In our applications considered in Section 3, the survey data are obtained from stratified
simple random samples without replacement, with strata defined by enterprise characteristics
such as economic activity and size class. The Rao–Wu–Yue bootstrap was developed for
general stratified multistage designs and can be directly applied to stratified simple random
sampling.Where the paper says it
Where the paper says it
Tam, S.-M. and Holmberg, A. p. 1 This setting makes classical calibration or model-assisted estimation methods inapplicable
in their standard form. From the perspective of non-probability sampling, our setting is best understood in terms
of the four-type classification summarized by Tam and Holmberg (2020). This setting makes classical calibration or model-assisted estimation methods inapplicable
in their standard form. From the perspective of non-probability sampling, our setting is best understood in terms
of the four-type classification summarized by Tam and Holmberg (2020). These totals and
their variance estimates are then used in a Fay–Herriot area-level model with exactly
known covariates to produce empirical best linear unbiased predictors. The approach
is demonstrated in three enterprise survey settings from official statistics by integrating probability sample data with (i) administrative records, (ii) a cut-off data source,
and (iii) web-scraped online information. These totals and
their variance estimates are then used in a Fay–Herriot area-level model with exactly
known covariates to produce empirical best linear unbiased predictors. The approach
is demonstrated in three enterprise survey settings from official statistics by integrating probability sample data with (i) administrative records, (ii) a cut-off data source,
and (iii) web-scraped online information.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Wu, C. and Sitter, R. p. 1 This contrast highlights
a key advantage of our framework: valid inference without ignorability assumptions on the
selection mechanism. We develop a methodology that combines probability sample data with incomplete auxiliary information, following the model-calibration framework of Wu and Sitter (2001) and
adapting ideas from Kim and Tam (2021) on data integration. This contrast highlights
a key advantage of our framework: valid inference without ignorability assumptions on the
selection mechanism. We develop a methodology that combines probability sample data with incomplete auxiliary information, following the model-calibration framework of Wu and Sitter (2001) and
adapting ideas from Kim and Tam (2021) on data integration.Where the paper says it
Where the paper says it
Ybarra, L. M. R. and Lohr, S. p. 2 The resulting
calibrated estimators can then serve as inputs to SAE models, further reducing estimation
uncertainty. Second, we demonstrate through applications to Lithuanian official statistics
that the proposed method outperforms alternative approaches, including a modification of
the FH model that incorporates covariate measurement error (Ybarra and Lohr, 2008). The resulting
calibrated estimators can then serve as inputs to SAE models, further reducing estimation
uncertainty. Second, we demonstrate through applications to Lithuanian official statistics
that the proposed method outperforms alternative approaches, including a modification of
the FH model that incorporates covariate measurement error (Ybarra and Lohr, 2008).Where the paper says it
Where the paper says it
--- Page 1 ---
Small area estimation using incomplete auxiliary information proposes a twostep approach that uses design-based model calibration with non-probability data to obtain approximately design-unbiased domain totals, which are then used in a Fay–Herriot area-level model to produce empirical best linear unbiased predictors
Battese, G. E., Harter, R. M., and Fuller, W. A. (1988). An error-components model for prediction of county crop areas using survey and satellite data
Booth, J. G., Butler, R. W., and Hall, P. (1994). Bootstrap methods for finite populations
Chen, Y., Li, P., and Wu, C. (2020). Doubly robust inference with nonprobability survey samples
Deville, J.-C. and Särndal, C.-E. (1992). Calibration estimators in survey sampling
Elliott, M. and Valliant, R. (2017). Inference for nonprobability samples
Fay, R. E. and Herriot, R. A. (1979). Estimates of income for small places: an application of James-Stein procedures to census data
Harmening, S., Kreutzmann, A.-K., Schmidt, S., Salvati, N., and Schmid, T. (2023). A framework for producing small area estimates based on area-level models in R
Jiang, J., Lahiri, P., and Wan, S.-M. (2002). A unified jackknife theory for empirical best prediction with M-estimation
Kim, J., Park, S., Chen, Y., and Wu, C. (2021). Combining non-probability and probability survey samples through mass imputation
Kim, J.-K. and Tam, S.-M. (2021). Data integration by combining big data and survey sample data for finite population inference
Liu, A.-C., Scholtus, S., and de Waal, T. (2023). Correcting selection bias in big data by pseudo-weighting
Medous, E., Goga, C., Ruiz-Gazen, A., Beaumont, J.-F., Dessertaine, A., and Puech, P. (2024). QR prediction for statistical data integration
Rao, J. N. K. and Molina, I. (2015). Small Area Estimation
Rao, J. N. K., Wu, C., and Yue, K. (1992). Some recent work on resampling methods for complex surveys
Tam, S.-M. and Holmberg, A. (2020). New data sources for official statistics – a game changer for survey statisticians? The Survey Statistician
Wu, C. and Sitter, R. R. (2001). A model-calibration approach to using complete auxiliary information from survey data
Ybarra, L. M. R. and Lohr, S. L. (2008).
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Model calibration for domain totals can be used to obtain
approximately design-unbiased domain totals
by fitting a unit-level measurement error working model onthe linked overlap between the probability survey and the external source.
This allows AI systems to produce more reliable estimates of small area totals when auxiliary information is incomplete or noisy, as demonstrated in applications like monthly turnover. -
The framework enables
valid inference without ignorability assumptions on the selection mechanism
of non-probability samples, contrasting with traditional methods thatrely on an unverifiable missing-at-random (MAR) assumption.
This allows AI systems to leverage large, non-probability data sources for estimation without needing to model the complex selection process of those sources. -
The approach allows for
substantially more general relationships between the observed proxies and the study variables
than linear models, accommodatinglinear, nonlinear, or nonparametric relationships
in the measurement error model. This enables AI systems to integrate diverse auxiliary information types into estimation tasks where simple linear assumptions are insufficient. -
The area-level modeling step can be implemented using a Fay Herriot model on log-transformed domain totals, which produces
empirical best linear unbiased predictors (EBLUPs)
thatborrow strength across domains using covariates based on completely known auxiliary data.
This allows AI systems to generate predictions for unknown areas by leveraging information from areas with known auxiliary data. -
The Ybarra-Lohr formulation extends the model to include an estimated auxiliary total, allowing the calculation of a shrinkage factor gamma hat m that accounts for measurement error uncertainty in the proxy total. This provides a
model-based improvement over the direct estimators
by incorporating information about how much uncertainty exists in the auxiliary information itself.
Abstract
Auxiliary information is increasingly available from administrative and other data sources, but it is often incomplete and of non-probability origin. We propose a two-step small area estimation approach in which the first step relies on design-based model calibration and exploits a large non-probability sample providing a noisy proxy of the study variable for only part of the population. A unit-level measurement-error working model is fitted on the linked overlap between the probability survey and the non-probability sample, and its predictions are incorporated through domain-specific model-calibration constraints to obtain approximately design-unbiased domain totals. These totals and their variance estimates are then used in a Fay-Herriot area-level model with exactly known covariates to produce empirical best linear unbiased predictors. The approach is demonstrated in three enterprise survey settings from official statistics by integrating probability sample data with (i) administrative records, (ii) a cut-off sample, and (iii) web-scraped online information. Empirical comparisons show consistent improvements in domain-level precision over direct estimation and over a Fay-Herriot benchmark that directly incorporates the proxy information as an error-prone covariate. These gains are achieved without modeling the mechanism that generates the non-probability sample: the realized sample is used as fixed auxiliary information, while the working prediction function is estimated from the probability sample under the sampling design.
Related papers
- Neural Estimation of Summary Statistics for Simulation-Based Inference with Non-Asymptotic Guarantees
- Spectral divide-and-conquer MCMC for long stationary time series
- Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE
- Personalized Federated Vector Autoregression with Personalization Diversity
- Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes
- Covariance-Calibrated Prediction-Powered Inference